In 2025 (seems a lifetime ago), I was transitioning to building using specs and plans. Driving driving everything myself and keeping things on track.
I was directing the flow. I had automated the verifications but I was triggering the learning. I was aware of context and degradation.
I was already working on Foundations of AI Coding.
But as much as I had automated, I was stil slow.
I'd love to say the solution was simple :) It wasn't but it was certainly the right solution - building an agent harness.
My current harness has seen quite a number of iterations to be able to route work to the right model and work well with Claude, Codex, Kimi and a few more.
And I'm building sooo much faster now.
I've been working on "Build your Agentic Harness" for a couple of months. The end-goal is for you to have a full, working harness by the end.
It's on the roadmap and of course it's going to be in All Access:
Getting here is a bit of a story :)
Both OpenAI and Anthopic published their now seminal pieces on building with an agentic harness.
Initially I remember thinking that would be overkill for what I was building. And likely too complex.
I re-read both articles and continued to think about what an implementation would look like.
Maybe it wasn't necessary too complicated. And it didn't have to mean lower quality in the codebase.
And it would definitely solve some of my problems like having to go down the stairs to check if the agent was done or stuck :)
There are definitely levels to this.
Hopefully the new material will help you move a lot faster.
Talk to you soon. Cheers,
Eugen.
|
|